feat: add end-to-end agent diagnostics command

[GautamKumarOffical]

Summary

Adds agn diagnose <agent-name> command that runs structured health checks to determine if an agent is truly usable end-to-end.

Problem

Users currently need to manually combine multiple status surfaces (agn status, /v1/health, /v1/discover, heartbeat logs) to answer the question: "Can this agent actually receive and complete work right now?" This makes debugging "agent is running but not responding" problems difficult.

Solution

New agn diagnose command that runs a comprehensive set of checks:

1. Daemon liveness: Is the daemon process running?
2. Adapter state: Is the adapter marked as running?
3. Workspace health: Is the workspace backend reachable?
4. Workspace presence: Is the agent online with fresh heartbeats?
5. Session validity: Can the agent heartbeat successfully?
6. LLM connectivity: Can the configured LLM provider respond?

Output Format

Diagnosing agent: my-agent (claude)

  ✓ daemon                 PID 12345
  ✓ adapter_state          running
  ✓ workspace_health       reachable
  ✓ workspace_presence     online (12s ago)
  ✓ session                valid
  ✓ llm                    claude-3-5-sonnet-20241022

Result: ALL CHECKS PASSED — agent is ready

Each check shows ok/warn/error status, making it easy to identify exactly where an agent's problems lie.

Changes

• packages/agent-connector/src/cli.js:
  • Add cmdDiagnose() function with structured health checks
  • Add 'diagnose' to help text and command handler

Usage

agn diagnose <agent-name>

Fixes #485

[vercel]

Someone is attempting to deploy a commit to the Raphael's projects Team on Vercel.

A member of the Team first needs to authorize it.

[zomux]

Closing this one with thanks. The problem it targets — needing to stitch together agn status, /v1/health, /v1/discover and heartbeat logs to tell whether an agent can really take work — has since been covered on develop by the live agent probe and smoke-test path (agn probe / the launcher's agent smoke tests, shipped 2026-08-21), which exercises the round trip end to end and prints failure guidance. This branch is three months behind cli.js and would need a rewrite against that rather than a rebase. If there's a specific diagnostic the probe still doesn't surface, a small PR adding it to the probe would be very welcome.